WEBVTT

00:00:00.000 --> 00:00:03.277
I think AI may exceed the sum of human

00:00:03.360 --> 00:00:06.957
intelligence in about in around 5 years.

00:00:07.040 --> 00:00:10.077
>> In 5 years?
>> Roughly 5 years is my guess.

00:00:10.160 --> 00:00:15.277
Um So there really
won't be anything that AI

00:00:15.360 --> 00:00:20.117
can can't do better than humans
apart from being human perhaps.

00:00:20.200 --> 00:00:25.320
>> In a more prosaic level
what will life be like?

00:00:25.800 --> 00:00:28.077
>> The most likely outcome

00:00:28.160 --> 00:00:31.597
is an age of amazing abundance

00:00:31.680 --> 00:00:35.837
uh where anyone can have
anything they can think

00:00:35.920 --> 00:00:38.877
of. This is this is this may sound

00:00:38.960 --> 00:00:43.117
preposterous but well here we are in 2026.

00:00:43.200 --> 00:00:45.277
Let's see where we stand in 2036. Now of

00:00:45.360 --> 00:00:50.397
course there are things that could
uh derail this like a a massive global

00:00:50.480 --> 00:00:54.917
thermonuclear war or something like that.
Um but assuming that there's not World War

00:00:55.000 --> 00:00:59.637
III uh I think I think
we're headed for an age of

00:00:59.720 --> 00:01:03.397
amazing abundance.
So this is I guess a message of

00:01:03.480 --> 00:01:07.557
of optimism and uh
excitement about the future.

00:01:07.640 --> 00:01:09.517
>> You're sure that humans will no longer

00:01:09.600 --> 00:01:11.837
be in control in 10 years time?

00:01:11.920 --> 00:01:15.797
>> I think it is unlikely
um in the same way that

00:01:15.880 --> 00:01:20.397
if you had a vastly more like a a

00:01:20.480 --> 00:01:24.037
you know the

00:01:24.120 --> 00:01:25.917
like like if if if the difference in the

00:01:26.000 --> 00:01:29.917
intelligence between AI and humans is

00:01:30.000 --> 00:01:34.237
vastly greater than the difference
in intelligence between AI and chimpanzees

00:01:34.320 --> 00:01:36.237
it's hard to imagine that the

00:01:36.320 --> 00:01:38.757
chimpanzees would be in charge.

00:01:38.840 --> 00:01:42.277
>> I'm really stuck by this. >> um

00:01:42.360 --> 00:01:47.037
an evolved form of chimpanzee.
So it wasn't that long ago that we were

00:01:47.120 --> 00:01:52.437
swinging through the trees eating
bananas um in the jungle. And we still

00:01:52.520 --> 00:01:55.397
eat bananas. >> You said
there was a 10 to 20% chance of

00:01:55.480 --> 00:01:58.037
killer robots wiping out humanity. So,

00:01:58.120 --> 00:02:02.437
>> you sounded worried then. Now you seem
very relaxed about it. Have you changed

00:02:02.520 --> 00:02:07.077
your view? >> I still think there's risk
associated with AI and robots. It's not

00:02:07.160 --> 00:02:08.397
zero.

00:02:08.480 --> 00:02:12.597
Um I I guess I've come to

00:02:12.680 --> 00:02:17.877
my my I sort of philosophical
conclusion is to look on the bright side.

00:02:17.960 --> 00:02:21.397
Um I like I I can't see any way to really

00:02:21.480 --> 00:02:26.637
stop this incredible
momentum of AI and robots.

00:02:26.720 --> 00:02:29.717
Um and and at times I sort of think, well,

00:02:29.800 --> 00:02:33.717
perhaps, you know, we're even if there
was a stop button, we probably shouldn't

00:02:33.800 --> 00:02:37.797
press it uh because

00:02:37.880 --> 00:02:42.877
I cuz the most likely
outcome is incredible

00:02:42.960 --> 00:02:46.637
abundance for for all.
>> That's what I thought, actually. It

00:02:46.720 --> 00:02:51.037
struck me that now you seem
to have decided it's inevitable.

00:02:51.120 --> 00:02:54.677
Let's hope for the best
and go for the ride.

00:02:54.760 --> 00:02:58.277
>> Yeah. I mean yes, pretty much.

00:02:58.360 --> 00:03:03.317
Let's uh let's enjoy the ride as as I
think my philosophy at this point. Cuz

00:03:03.400 --> 00:03:06.637
I mean, honestly, this this it's
it's there seems to be just an

00:03:06.720 --> 00:03:12.317
inexorable uh progress
here in in AI and and in

00:03:12.400 --> 00:03:15.837
robots that um >> But you

00:03:15.920 --> 00:03:19.637
>> you know, even if I wanted
to stop it, I I don't I couldn't.

00:03:19.720 --> 00:03:21.677
>> Do you think there's still a 10 to 20%

00:03:21.760 --> 00:03:23.957
chance that killer robots

00:03:24.040 --> 00:03:27.157
mean the end of humanity?

00:03:27.240 --> 00:03:30.037
That's a high bar to as they say.

00:03:30.120 --> 00:03:33.837
>> Well, in long run
we're all dead, of course.

00:03:33.920 --> 00:03:38.797
Um you know, um So, I mean,
I I I so >> I mean >> the let let's

00:03:38.880 --> 00:03:41.917
according to the standard
model of physics, the end

00:03:42.000 --> 00:03:44.677
of the universe is um

00:03:44.760 --> 00:03:47.117
a long, boring

00:03:47.200 --> 00:03:50.997
sort of thin, gray soup of

00:03:51.080 --> 00:03:54.157
uh gradual cooling.

00:03:54.240 --> 00:03:58.557
Um I mean, it is sort of the heat death
of the universe is called. So, now if if

00:03:58.640 --> 00:04:03.117
the heat death of the universe is in
fact the outcome for our reality, then

00:04:03.200 --> 00:04:05.557
it really is all about the journey.

00:04:05.640 --> 00:04:07.117
Because the destination is terrible.

00:04:07.200 --> 00:04:09.637
>> years is not [laughter] very Yeah,
yeah, I get that, but that could

00:04:09.720 --> 00:04:12.597
be a long way ahead.
>> long very very very very long way.

00:04:12.680 --> 00:04:15.317
>> If it's 10 to 20%
in 10 years, you would

00:04:15.400 --> 00:04:18.197
you go into one of your rockets if you
thought there was a 10 to 20% chance of

00:04:18.280 --> 00:04:20.917
it blowing up and killing you?

00:04:21.000 --> 00:04:25.637
>> Yes, but I I would >> [laughter]
>> But let's say there's you can't do

00:04:25.720 --> 00:04:29.317
anything about it. >> Okay. >> But I
guess the only thing we can try to

00:04:29.400 --> 00:04:33.317
do is minimize the probability
of the bad outcome.

00:04:33.400 --> 00:04:36.677
Um but trying to stop the intelligence in

00:04:36.760 --> 00:04:38.557
robots

00:04:38.640 --> 00:04:40.837
um

00:04:40.920 --> 00:04:44.317
I mean, for the longest time I

00:04:44.400 --> 00:04:48.237
declined to participate
in AI or was, you know,

00:04:48.320 --> 00:04:52.717
you know, created OpenAI as essentially
a counterweight to to Google

00:04:52.800 --> 00:04:57.317
cuz at the time they were had
more or less a monopoly on AI. In

00:04:57.400 --> 00:05:01.437
creating OpenAI and then having
Anthropic spin out of that and now

00:05:01.520 --> 00:05:03.637
currently Anthropic's leader in AI.

00:05:03.720 --> 00:05:06.917
Um I mean, these

00:05:07.000 --> 00:05:10.077
these actions have actually resulted in

00:05:10.160 --> 00:05:13.197
knock-on effects that accelerated AI

00:05:13.280 --> 00:05:15.797
which wasn't really my intention.

00:05:15.880 --> 00:05:19.197
So, it just seems like all roads lead to

00:05:19.280 --> 00:05:22.357
acceleration of AI.

00:05:22.440 --> 00:05:27.677
So, then I'm like, okay, well,
you can just sort of be sad about it or

00:05:27.760 --> 00:05:30.317
join the club, I suppose. >> Well,
you've certainly joined the club.

00:05:30.400 --> 00:05:33.117
>> Yeah. >> I mean, you've >> [laughter]
>> you've definitely joined the club.

00:05:33.200 --> 00:05:34.797
>> If you can't beat them,
join them, you know.

00:05:34.880 --> 00:05:38.357
>> A number of the other leaders of of

00:05:38.440 --> 00:05:42.237
model makers have been pushing
for different ways to try and minimize the

00:05:42.320 --> 00:05:44.717
risks. Um you know, just last week Demis

00:05:44.800 --> 00:05:47.797
Hassabis came out with a proposal for a

00:05:47.880 --> 00:05:50.037
private public regulator. Do you think

00:05:50.120 --> 00:05:53.397
you guys should be working more actively

00:05:53.480 --> 00:05:57.197
to get something in place to control the
risk, or do you think that's impossible,

00:05:57.280 --> 00:06:01.797
too? >> Yeah. My recommendation
to Demis cuz we I talked to him for

00:06:01.880 --> 00:06:06.317
a few hours before he published
his his piece was that um

00:06:06.400 --> 00:06:10.197
the most immediate thing that we could
do is to have the leading AI companies

00:06:10.280 --> 00:06:14.957
um at least just meet uh
or have a some sort of call

00:06:15.040 --> 00:06:19.397
once every few weeks um and just

00:06:19.480 --> 00:06:23.757
uh discuss say any safety
and security issues.

00:06:23.840 --> 00:06:26.917
>> But why why haven't you done that yet?
I mean, you don't all like each other very

00:06:27.000 --> 00:06:30.957
much. Yeah, but you don't you're not
all I mean, you're also all competing,

00:06:31.040 --> 00:06:32.837
right? You all want to be first.

00:06:32.920 --> 00:06:35.917
>> but that's that's where I think um good

00:06:36.000 --> 00:06:39.997
the competitors can keep each
keep each other honest. So, meaning

00:06:40.080 --> 00:06:43.917
meaning like if um like I think it's
quite difficult for someone in the

00:06:44.000 --> 00:06:46.917
government who doesn't have a deep
technical understanding and isn't

00:06:47.000 --> 00:06:51.797
driving the frontier of of AI to know
whether something should be released or

00:06:51.880 --> 00:06:55.997
not. Um however, the competitors

00:06:56.080 --> 00:06:58.277
I think what can, you know, if given

00:06:58.360 --> 00:07:00.317
sort of a week or two to review a new

00:07:00.400 --> 00:07:04.277
model um can highlight issues and uh

00:07:04.360 --> 00:07:07.677
and have an incentive to keep the all
all the competitors have an incentive to

00:07:07.760 --> 00:07:10.677
keep the others honest. >> So,
you'd have a system where the model

00:07:10.760 --> 00:07:13.757
makers would check each other's work.
If there was something that worried them,

00:07:13.840 --> 00:07:15.037
you'd tell the government. >> Yeah.

00:07:15.120 --> 00:07:17.317
>> The government would then take
>> If if there's something else was

00:07:17.400 --> 00:07:20.797
worrying them and that company was not
doing anything to address that risk,

00:07:20.880 --> 00:07:24.877
then that would be the the moment
for government to step in

00:07:24.960 --> 00:07:27.957
and take action, which is literally
what just what occurred with my model.

00:07:28.040 --> 00:07:31.120
>> this all have to happen?

00:07:31.480 --> 00:07:33.437
>> Well, I think it'd be good to do that.

00:07:33.520 --> 00:07:37.997
I think immediately, really.
>> Yeah, that's what people keep telling

00:07:38.080 --> 00:07:41.997
me. We we have to get this some kind
of system in the next 6 months because of

00:07:42.080 --> 00:07:45.437
the pace at which
the models are improving.

00:07:45.520 --> 00:07:47.757
>> Yeah, 6 months is a long time.

00:07:47.840 --> 00:07:51.877
Um >> So, you think shorter than that?

00:07:51.960 --> 00:07:55.157
>> Well, I mean there's there's so many AI

00:07:55.240 --> 00:07:59.277
breakthrough announcements.
Sometimes multiple per day.

00:07:59.360 --> 00:08:00.877
I mean, how many were there last week?

00:08:00.960 --> 00:08:02.557
>> A lot. >> I lost track, you know.

00:08:02.640 --> 00:08:06.637
>> Yeah, a lot. >> Um

00:08:06.720 --> 00:08:10.317
>> So, so you you basically, you know,
whether it's Demis's proposal or others,

00:08:10.400 --> 00:08:13.237
you need to get together and get
this set up kind of now. It sounds like

00:08:13.320 --> 00:08:14.557
you're telling me.

00:08:14.640 --> 00:08:17.557
>> Yeah, I I I would recommend that we at

00:08:17.640 --> 00:08:21.797
least have some sort
of informal um you know,

00:08:21.880 --> 00:08:26.357
weekly or bi-weekly um call and

00:08:26.440 --> 00:08:29.917
and and that like so just
that that there's that that

00:08:30.000 --> 00:08:31.637
um there's maybe

00:08:31.720 --> 00:08:35.197
a week or 2 weeks of of early access by

00:08:35.280 --> 00:08:39.597
competitors. And and this is why I think
the incentives work is because if if,

00:08:39.680 --> 00:08:42.957
you know, competing companies
I think are likely to

00:08:43.040 --> 00:08:46.677
understand if something is a security
risk. And they're they're not going to

00:08:46.760 --> 00:08:50.757
be shy about highlighting that that
their their competitors' models should

00:08:50.840 --> 00:08:54.677
be delayed in release.
>> You guys don't much like each other and

00:08:54.760 --> 00:08:57.677
you don't much trust each other. You're
insulting each other on on social media

00:08:57.760 --> 00:09:01.997
constantly. So, do you do you think you
can do that with people that you dislike

00:09:02.080 --> 00:09:05.237
that that where there seems to be such
kind of personal vitriol between some of

00:09:05.320 --> 00:09:10.957
these leaders? >> Um Yes, well,
I'm not I'm not a fan of Sam

00:09:11.040 --> 00:09:14.557
Altman because if you started
a nonprofit that was meant to be an

00:09:14.640 --> 00:09:19.917
open-source AI company and um

00:09:20.000 --> 00:09:22.637
you know, owned by the
world and it somehow got

00:09:22.720 --> 00:09:26.437
turned into an $800
billion for-profit company

00:09:26.520 --> 00:09:29.917
with closed source, I think you'd
be like, "Well, wait a second. That's

00:09:30.000 --> 00:09:34.477
the exact opposite of what I donated
the money for." That that's my issue. I

00:09:34.560 --> 00:09:37.237
think it's a legitimate one. Um

00:09:37.320 --> 00:09:42.037
And then Dario I think also is not a fan
of Sam Altman because he felt in fact

00:09:42.120 --> 00:09:46.917
the reason the Anthropic team left OpenAI
is because they didn't trust Sam Altman.

00:09:47.000 --> 00:09:50.957
Otherwise Anthropic wouldn't exist.
They would still be at OpenAI.

00:09:51.040 --> 00:09:56.877
Um I I think Dario is a very principled
person and he and he cares about things

00:09:56.960 --> 00:09:59.877
a lot. Uh he cares about
the future of of the

00:09:59.960 --> 00:10:04.477
world and I I think everyone I've
met at Anthropic so far has been

00:10:04.560 --> 00:10:07.637
I'm on Like no one set off my evil
detector. You know, they always seem

00:10:07.720 --> 00:10:09.957
seem like they're well-intentioned.

00:10:10.040 --> 00:10:14.077
Um and uh and the road
to hell is I think mostly

00:10:14.160 --> 00:10:17.797
paved with bad intentions. There are
a few well-intentioned paving stones in

00:10:17.880 --> 00:10:20.597
there, so we we don't
want to be complacent.

00:10:20.680 --> 00:10:25.877
Um but uh you know, at the end of the day
if we have to talk, we'll talk. I mean

00:10:25.960 --> 00:10:30.880
set aside our personal differences
for the good of the world type of thing.